A method and system for early warning of leakage faults in hydrogen energy storage devices based on deep learning

Abstract

This application provides a method and system for early warning of leakage faults in hydrogen energy storage equipment based on deep learning. The method includes: real-time collection of multi-source sensor data from the hydrogen energy storage equipment, construction of a multi-dimensional feature matrix, and establishment of a nonlinear mapping between pressure fluctuations and gas concentration using a bidirectional interactive network. Spatial modeling of temperature gradients and vibration spectra is performed to generate a fused feature tensor. Subsequently, the fused feature tensor is input into a spatiotemporal inference model to predict equipment leakage and energy loss, outputting a leakage probability cloud map and an energy loss vector. Based on the energy loss vector, regions of entropy abrupt changes are identified. When a high-density leakage region overlaps with an entropy abrupt change region within a continuous detection cycle, a multi-level early warning signal is triggered, generating a leakage source heat map and a pressure balance parameter set. This activates a self-regulating system to dynamically correct valve openings and update network weights, forming a closed-loop control link. This application improves the accuracy and response speed of early warning for leakage faults in hydrogen energy storage equipment.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a method and system for early warning of leakage faults in hydrogen energy storage equipment based on deep learning. Background Technology

[0002] Hydrogen energy storage equipment has important applications in the field of new energy, but during its operation, leakage malfunctions may lead to energy loss or even safety accidents.

[0003] Currently, there are leak detection solutions based on single-sensor data analysis and machine learning algorithms. For example, support vector machines or random forest algorithms are used to classify pressure or gas concentration data to determine whether there is a risk of leakage. These solutions, by training historical data models, can identify leak characteristics to a certain extent and trigger early warning signals.

[0004] Existing solutions primarily rely on data from a single type of sensor, failing to fully leverage the complementarity of multi-source heterogeneous data, resulting in insufficient accuracy and robustness in leak detection. Furthermore, existing solutions lack detailed modeling of leak propagation paths and energy losses, making it impossible to locate and dynamically adjust leak sources, and thus failing to meet the real-time early warning and control requirements under complex operating conditions. Summary of the Invention

[0005] This application provides a method and system for early warning of leakage faults in hydrogen energy storage equipment based on deep learning, in order to solve the problems of poor early warning accuracy and slow response speed of leakage faults in hydrogen energy storage equipment in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for early warning of leakage faults in hydrogen energy storage devices based on deep learning, including:

[0007] Multi-source heterogeneous sensor data streams from hydrogen energy storage devices are collected in real time to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra.

[0008] The multidimensional feature matrix is ​​input into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network. The gradient field reconstruction algorithm is used to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor.

[0009] The fused feature tensor is input into the spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage equipment, and outputs a leakage probability cloud map and an energy loss vector.

[0010] Based on the energy loss vector, the entropy value mutation region is determined. When the high-density region in the leakage probability cloud map and the entropy value mutation region spatially overlap within a continuous detection cycle, a multi-level early warning signal is triggered and a leakage source thermal map and pressure balance parameter set are generated.

[0011] The self-regulating system is activated based on the leak source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

[0012] Optionally, the spatiotemporal reasoning model includes a multi-layer perception constraint module, an implicit state transition model, a gas diffusion physical constraint module, and a probability density clustering module connected in sequence.

[0013] The step of inputting the fused feature tensor into the spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage equipment, and outputting a leakage probability cloud map and an energy loss vector, includes:

[0014] The internal structural connection relationship of the hydrogen energy storage device is analyzed by a multi-layer sensing constraint module to generate an energy transfer path topology map.

[0015] Based on the energy transfer path topology, the energy flow field feature matrix is ​​tracked by a dynamic recursive operator in the implicit state transition model. The dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform.

[0016] The energy flow field feature matrix is ​​input into the gas diffusion physics constraint module, and a three-dimensional spatial grid coordinate system is constructed by combining real-time environmental parameters.

[0017] The diffusion path of leaked gas in a three-dimensional spatial grid coordinate system is simulated by the mass conservation equation and the momentum transfer theorem, and the probability distribution of the diffusion path of leaked gas is generated.

[0018] A fusion leakage situation map is generated based on the probability distribution of the diffusion path;

[0019] Based on the fused leakage situation map, a leakage probability cloud map is generated through the probability density clustering module, and an energy loss vector is constructed according to the energy flow field feature matrix.

[0020] Optionally, the step of tracking the energy flow field feature matrix based on the energy transfer path topology map using a dynamic recursive operator in the implicit state transition model, wherein the dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform, includes:

[0021] The temperature gradient distribution field is decomposed into heat conduction and thermal stress components by a preset thermoelastic coupling equation, and a mechanical vibration influence factor matrix is ​​generated based on the frequency domain energy distribution of the vibration spectrum waveform.

[0022] A dual-channel convolution fusion operation is performed on the heat conduction component and the mechanical vibration influence factor matrix to generate a temperature vibration coupling coefficient matrix. The dimension of the coupling coefficient matrix is ​​consistent with the number of nodes in the energy transfer path topology graph.

[0023] A physically constrained state transition calculation is performed on the energy transfer path topology, and the energy flow field feature matrix is ​​iteratively updated. In each iteration, the dynamic recursive operator dynamically adjusts the state transition step size according to the spectral radius of the temperature vibration coupling coefficient matrix.

[0024] Optionally, before performing physically constrained state transition calculations on the energy transfer path topology graph, the method further includes:

[0025] Frequency domain energy spectrum analysis was performed on the pressure fluctuation sequence to obtain the analysis results;

[0026] Energy mutation indexes in different frequency bands are extracted from the analysis results using a bandpass filter window to construct a pressure frequency domain mutation feature vector.

[0027] And calculate the spatial gradient direction weighting coefficients based on the preset leakage diffusion direction constraints;

[0028] The pressure frequency domain abrupt change feature vector and the spatial gradient direction weight coefficient are input into the dual-channel attention gating module to generate the initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module;

[0029] The initial value of the time-varying gain factor is dynamically normalized using the leakage risk index threshold surface to generate a normalized time-varying gain factor.

[0030] The normalized time-varying gain factor is multiplied by the temperature vibration coupling coefficient matrix to generate a dynamic weight parameter matrix.

[0031] By embedding the dynamic weight parameter matrix as a physical constraint into the initial state transition equation, a state transition equation with physical constraints is obtained.

[0032] Optionally, the step of inputting the pressure frequency domain abrupt change feature vector and the spatial gradient direction weighting coefficient into the dual-channel attention gating module to generate an initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module includes:

[0033] Perform a multi-scale decomposition operation on the pressure frequency domain abrupt change feature vector to obtain multi-scale decomposition information;

[0034] Wavelet packet transform is performed on the multi-scale decomposition information to extract the energy mutation intensity coefficients of different frequency bands;

[0035] Based on the energy mutation intensity coefficients of different frequency bands, construct the pressure frequency domain energy distribution tensor;

[0036] The spatial gradient direction weight coefficients are subjected to directional enhancement processing to obtain the processed spatial gradient direction weight coefficients.

[0037] Based on the processed spatial gradient direction weight coefficients, an anisotropic diffusion filtering algorithm is used to generate a gradient direction consistency matrix.

[0038] The directional weight correction factor is calculated based on the gradient direction consistency matrix and the preset leakage diffusion direction constraint.

[0039] The pressure frequency domain energy distribution tensor and the direction weight correction factor are input into the dual-channel attention gating module to establish a nonlinear correlation mapping between the pressure frequency domain and the spatial gradient through the cross attention module in the dual-channel attention gating module, thereby generating an initial attention weight matrix.

[0040] A nonlinear activation operation is performed on the initial attention weight matrix to generate initial values ​​for the time-varying gain factor.

[0041] Optionally, the step of inputting the multidimensional feature matrix into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross-attention module in the bidirectional interactive network, and using a gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor, includes:

[0042] Perform time-frequency joint analysis on the pressure fluctuation sequence to generate a pressure fluctuation time-frequency feature matrix;

[0043] Perform spatial multi-resolution analysis on the gas concentration spectrum to generate a multi-scale feature map of gas concentration;

[0044] The pressure fluctuation time-frequency feature matrix and the gas concentration multi-scale feature map are input into the cross-attention module to establish the correlation mapping between the pressure frequency domain features and the gas concentration spatial features through the multi-head attention module, thereby generating a pressure-concentration interaction feature matrix.

[0045] A temperature gradient propagation path map is generated based on the temperature gradient distribution field.

[0046] The intrinsic mode function components of the vibration spectrum waveform are extracted using the empirical mode decomposition algorithm, and a vibration spectrum energy feature vector is generated based on the intrinsic mode function components.

[0047] Based on the temperature gradient propagation path diagram and the vibration spectrum energy feature vector, a temperature vibration interaction feature matrix is ​​generated;

[0048] The pressure-concentration interaction feature matrix and the temperature-vibration interaction feature matrix are joined by tensor concatenation to generate a fused feature tensor.

[0049] Optionally, the step of determining the entropy abrupt change region based on the energy loss vector, when the high-density region in the leakage probability cloud map and the entropy abrupt change region spatially overlap within a continuous detection period, triggers a multi-level early warning signal and generates a leakage source thermal map and pressure balance parameter set, including:

[0050] Perform a multi-dimensional decomposition operation on the energy loss vector to generate an energy loss feature matrix;

[0051] Abnormal energy entropy values ​​are identified based on the energy loss feature matrix to generate entropy value mutation regions;

[0052] Spatial clustering analysis is performed on the leakage probability cloud map to generate high-density regions;

[0053] When the high-density region and the entropy value mutation region spatially overlap for at least two consecutive detection cycles, a multi-level early warning signal is triggered.

[0054] Generate a leakage source thermal map corresponding to the high-density area. Based on the leakage source thermal map and the preset pressure balance equation, calculate the pressure compensation parameters of multiple adjacent hydrogen energy storage devices to generate a pressure balance parameter set.

[0055] Secondly, embodiments of this application provide a deep learning-based early warning system for leaks in hydrogen energy storage devices, comprising:

[0056] The collection module is used to collect multi-source heterogeneous sensor data streams from hydrogen energy storage devices in real time to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra.

[0057] The input module is used to input the multidimensional feature matrix into the bidirectional interactive network, so as to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network, and to use the gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor.

[0058] The output module is used to input the fused feature tensor into the spatiotemporal inference model, so as to use the spatiotemporal inference model to predict the leakage and energy loss of hydrogen energy storage equipment, and output the leakage probability cloud map and energy loss vector.

[0059] The triggering module is used to determine the entropy change region based on the energy loss vector. When the high-density region in the leakage probability cloud map and the entropy change region spatially overlap within a continuous detection period, it triggers a multi-level early warning signal and generates a leakage source thermal map and pressure balance parameter set.

[0060] The activation module is used to activate the self-regulating system based on the leakage source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

[0061] Thirdly, embodiments of this application provide a computing device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a method for early warning of leakage faults in hydrogen energy storage devices based on deep learning, as described in any of the first aspects.

[0062] Fourthly, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement the method for early warning of leakage faults in hydrogen energy storage devices based on deep learning as described in any one of the first aspects.

[0063] This application provides a method for early warning of leakage faults in hydrogen energy storage devices based on deep learning. The method includes: real-time collection of multi-source heterogeneous sensor data streams from the hydrogen energy storage device to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra. The multi-dimensional feature matrix is ​​input into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through a cross-attention module in the bidirectional interactive network. A gradient field reconstruction algorithm is then used to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor. The feature tensor is fused and input into a spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage devices, outputting a leakage probability cloud map and an energy loss vector. Based on the energy loss vector, entropy abrupt change regions are determined. When a high-density region in the leakage probability cloud map spatially overlaps with an entropy abrupt change region within a continuous detection period, a multi-level early warning signal is triggered, and a leakage source heat map and pressure balance parameter set are generated. A self-regulating system is activated based on the leakage source heat map and the pressure balance parameter set, enabling the self-regulating system to dynamically correct the valve opening gradient of adjacent hydrogen energy storage devices and update the node weights of the bidirectional interactive network to generate a closed-loop link.

[0064] The technical solution provided in this application has the following beneficial effects:

[0065] This application integrates multi-source data, including pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra, to construct a multi-dimensional feature matrix. This provides a comprehensive and high-dimensional data foundation for subsequent analysis, improving the comprehensiveness and accuracy of leak detection. A nonlinear mapping between pressure fluctuations and gas concentration is established using a cross-attention module. Combined with a gradient field reconstruction algorithm, spatial modeling of temperature gradients and vibration spectra is performed to generate a fused feature tensor, achieving deep fusion of multi-source data and enhancing feature representation capabilities. A spatiotemporal inference model is used to accurately predict leaks and energy losses, generating leak probability cloud maps and energy loss vectors, providing a basis for quantitative analysis and location of leak risks. By identifying regions of abrupt entropy changes and performing spatial overlap analysis with high-density regions in the leak probability cloud map, early warning of leak risks is achieved, triggering multi-level warning signals and improving the timeliness and accuracy of warnings. Based on the leak source thermal map and pressure balance parameter set, the valve opening gradient of adjacent equipment is dynamically corrected, and the node weights of the bidirectional interactive network are updated to form a closed-loop control link, realizing dynamic control and system optimization of leak risks.

[0066] Furthermore, this embodiment of the application analyzes the internal structural connection relationship of the hydrogen energy storage device through a multi-layer sensing constraint module to generate an energy transfer path topology map. Based on the topology map, the dynamic recursive operator in the implicit state transition model is used to track the energy flow field feature matrix, and the state transition step size is adaptively adjusted by combining the coupling effect of temperature gradient and vibration spectrum. The energy flow field feature matrix is ​​input into the gas diffusion physical constraint module to construct a three-dimensional spatial grid coordinate system. The leakage gas diffusion path is simulated through the mass conservation equation and momentum transfer theorem to generate a diffusion path probability distribution. A fused leakage situation map is generated based on the diffusion path probability distribution. Finally, a leakage probability cloud map is generated through the probability density clustering module, and an energy loss vector is constructed.

[0067] Furthermore, this step, through multi-layered sensing constraints and an implicit state transition model, accurately tracks the energy flow field characteristics. Combined with a gas diffusion physics constraint module, it achieves a three-dimensional simulation of the leakage gas diffusion path, generating a high-precision leakage probability cloud map and energy loss vector. This method improves the accuracy of leak location and the refinement of energy loss analysis, providing a reliable basis for quantitative assessment and dynamic control of leakage risks.

[0068] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 A flowchart illustrating a deep learning-based method for early warning of leaks in hydrogen energy storage devices, provided as an embodiment of this application.

[0071] Figure 2 A schematic diagram of a deep learning-based early warning system for hydrogen energy storage equipment leakage faults, provided as an embodiment of this application;

[0072] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0073] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0074] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] The research and development approach of this scheme is to collect multi-source heterogeneous sensor data from hydrogen energy storage devices in real time, construct a multi-dimensional feature matrix, and establish a nonlinear mapping between pressure fluctuations and gas concentration using the cross-attention module in a bidirectional interactive network. Simultaneously, a gradient field reconstruction algorithm is used to model the spatial propagation of temperature gradients and vibration spectra, generating a fused feature tensor. Next, the fused feature tensor is input into a spatiotemporal inference model to predict equipment leakage and energy loss, outputting a leakage probability cloud map and an energy loss vector. Based on the energy loss vector, regions of entropy abrupt changes are identified. When a high-density region in the leakage probability cloud map spatially overlaps with an entropy abrupt change region within a continuous detection cycle, multi-level early warning signals are triggered, and a leakage source thermal map and pressure balance parameter set are generated. Finally, based on the leakage source thermal map and pressure balance parameter set, a self-regulating system is activated to dynamically correct the valve opening gradient of adjacent devices and update the node weights of the bidirectional interactive network, forming a closed-loop control link to achieve accurate early warning and dynamic control of leakage risks.

[0077] Figure 1 A flowchart illustrating a deep learning-based method for early warning of leaks in hydrogen energy storage devices, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes:

[0078] Step 101: Collect multi-source heterogeneous sensor data streams from hydrogen energy storage devices in real time to construct a multi-dimensional feature matrix.

[0079] In this step, the sensor data stream includes a pressure fluctuation sequence, a temperature gradient distribution field, a vibration spectrum waveform, and a gas concentration spectrum. The pressure fluctuation sequence refers to continuous data showing the pressure change within the equipment over time, reflecting the stability of the equipment's operating state. The temperature gradient distribution field refers to the spatial distribution of temperature within the equipment, used to analyze heat conduction and thermal stress. The vibration spectrum waveform refers to the frequency domain distribution of the equipment's vibration signal, used to identify mechanical vibration characteristics. The gas concentration spectrum refers to the spatial distribution of gas concentration around the equipment, used to detect the diffusion of leaked gas. The multidimensional feature matrix integrates multi-source heterogeneous data into a high-dimensional matrix, facilitating subsequent analysis and modeling.

[0080] In this embodiment, various sensors (such as pressure sensors, temperature sensors, vibration sensors, and gas concentration sensors) installed on the hydrogen energy storage device are used to collect data such as pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra in real time. These data are aligned according to timestamps and spatial locations to construct a multi-dimensional feature matrix. Each row of the matrix represents multi-source data at a time point, and each column represents a feature dimension of sensor data.

[0081] For example, in a large hydrogen storage station, engineers installed various sensors at key locations on the hydrogen storage tanks to monitor changes in the hydrogen's state in real time. Using a dedicated data acquisition system, this raw data was transmitted to a central server and transformed into a multi-dimensional feature matrix, providing a foundation for further analysis.

[0082] Step 102: Input the multidimensional feature matrix into the bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network, and use the gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor.

[0083] In this step, the bidirectional interaction network refers to a deep learning model capable of simultaneously handling bidirectional interactions between multiple data sources. The cross-attention module is used to establish nonlinear mappings between different data features, such as the correlation between pressure fluctuations and gas concentration. The fusion feature tensor is a high-dimensional tensor generated by deeply fusing features from multiple data sources, used for subsequent prediction tasks.

[0084] In this embodiment, a multidimensional feature matrix is ​​input into a bidirectional interactive network, and the nonlinear mapping relationship between the pressure fluctuation sequence and the gas concentration spectrum is calculated through a cross-attention module. At the same time, a gradient field reconstruction algorithm is used to model the spatial propagation of the temperature gradient distribution field and the vibration spectrum waveform, and to analyze their coupling effect. Finally, the pressure-gas-concentration interaction features and the temperature-vibration interaction features are spliced ​​together to generate a fused feature tensor.

[0085] For example, based on the multidimensional feature matrix obtained in the previous step, the researchers used a bidirectional interactive network to conduct an in-depth analysis of the data. In particular, they identified the relationship between pressure fluctuations and abnormal gas concentrations at potential leakage risk points through a cross-attention module, and determined possible fault regions through a gradient field reconstruction algorithm, generating a fused feature tensor.

[0086] Step 103: Input the fused feature tensor into the spatiotemporal inference model to use the spatiotemporal inference model to predict the leakage and energy loss of hydrogen energy storage equipment, and output the leakage probability cloud map and energy loss vector.

[0087] In this step, the spatiotemporal inference model is a deep learning model that combines temporal and spatial features to predict leakage and energy loss. The leakage probability cloud map is a distribution map of the probability of leakage inside or around the equipment, used to locate areas of leakage risk. The energy loss vector is a vector representing the energy loss of the equipment, used to quantify the impact of leakage on equipment performance.

[0088] In this embodiment, the fusion feature tensor is input into the spatiotemporal inference model. The model predicts the probability of equipment leakage and energy loss by analyzing time and space features. The leakage probability cloud map is generated by probability density clustering to reflect the distribution of leakage risk. The energy loss vector is generated by analyzing the energy flow field feature matrix to quantify the energy loss caused by leakage.

[0089] For example, after analyzing the fused feature tensor using a spatiotemporal inference model, the system successfully predicted several high-risk leakage areas and calculated the corresponding energy loss vectors, providing an important basis for subsequent risk assessment.

[0090] Step 104: Determine the entropy change region based on the energy loss vector. When the high-density region in the leakage probability cloud map and the entropy change region spatially overlap within a continuous detection period, trigger a multi-level early warning signal and generate a leakage source thermal map and pressure balance parameter set.

[0091] In this step, the entropy abrupt change region refers to the area in the energy loss vector where the entropy value changes, indicating abnormal energy fluctuations. The high-density region refers to the area in the leakage probability cloud map where data points are densely distributed and exceed a preset density threshold. The distinction between high and low density is based on calculating the local density of data points using a spatial clustering algorithm and setting a density threshold. Areas above this threshold are considered high-density regions, indicating a higher probability of leakage; areas below the threshold are considered low-density regions, indicating a lower probability of leakage. A specific numerical example is as follows: setting a minimum number of points of 5 and a neighborhood radius of 0.5, if the number of data points per unit area in a region exceeds 10, it is considered a high-density region; if it is less than 5, it is considered a low-density region. For example, if a region has 12 data points per square meter, exceeding the threshold of 10, it is marked as a high-density region. The multi-level early warning signal is a tiered early warning mechanism triggered according to the severity of the leakage risk. Different levels of early warning signals correspond to different response measures to ensure timely and effective handling of leakage risks. Different warning levels correspond to different notification methods. Specifically, Level 1 warning (low risk) is indicated by voice prompts and pop-up notification boxes. For example, the system issues a voice prompt, "Potential leakage risk detected, please check the equipment," and a pop-up notification box appears on the monitoring interface. Level 2 warning (medium risk) is indicated by SMS notifications and light warnings. For example, the system sends an SMS to relevant personnel, "Medium leakage risk detected, please check the equipment immediately," and activates a yellow warning light at the equipment site. Level 3 warning (high risk) is indicated by emergency alarms and automatic system shutdown. For example, the system triggers an emergency alarm, sends an SMS, "High leakage risk detected, the system will automatically shut down soon," and automatically shuts down relevant equipment to prevent further leakage.

[0092] In this embodiment, entropy abrupt change regions are calculated based on energy loss vectors to identify abnormal energy fluctuations; spatial clustering analysis is performed on the leakage probability cloud map to determine high-density regions; when high-density regions and entropy abrupt change regions spatially overlap within a continuous detection cycle, multi-level early warning signals are triggered to generate a leakage source thermal map and pressure balance parameter set.

[0093] For example, during the detection process, the system detected increased energy loss in certain areas, indicating a potential leakage risk. Further analysis confirmed that these areas matched high-density zones on the leakage probability cloud map, prompting the activation of an early warning procedure and the generation of a detailed leak source heat map and adjustment recommendations.

[0094] Step 105: Activate the self-regulating system based on the leakage source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

[0095] In this step, a self-regulating system refers to a system capable of dynamically adjusting equipment operating parameters based on leakage risk. Valve opening gradient refers to the valve opening adjustment strategy of adjacent equipment valves, used to balance pressure distribution. A closed-loop link is a control link that achieves dynamic adjustment and optimization of the system through a feedback mechanism.

[0096] In this embodiment, the location of the leakage source is determined based on the leakage source heat map, and the valve opening gradient of adjacent equipment is calculated by combining the pressure balance parameter set; the self-regulating system dynamically adjusts the valve opening to balance the pressure distribution; at the same time, the node weights of the bidirectional interactive network are updated to optimize the model performance and form a closed-loop control link.

[0097] For example, upon receiving an early warning signal, the self-regulating system immediately takes action to adjust the valve opening of the relevant hydrogen storage tank, alleviating the local overpressure situation. At the same time, by continuously learning and adjusting network weights, it improves the safety and efficiency of the entire system.

[0098] This solution acquires multi-source heterogeneous data in real time, constructs a multi-dimensional feature matrix, and uses a bidirectional interactive network and spatiotemporal inference model to achieve accurate prediction of leakage and energy loss. Based on the overlapping analysis of entropy change regions and leakage probability cloud maps, it triggers multi-level early warning signals and generates a leakage source heat map and pressure balance parameter set. Finally, it dynamically adjusts the equipment operating parameters through a self-regulating system to form a closed-loop control link, thereby improving the accuracy of early warning and the efficiency of control of hydrogen energy storage equipment leakage faults.

[0099] To address the accuracy and precision issues in predicting leaks in hydrogen energy storage devices, in some embodiments, step 103: the spatiotemporal inference model includes a multilayer sensing constraint module, an implicit state transition model, a gas diffusion physical constraint module, and a probability density clustering module connected in sequence. The multilayer sensing constraint module refers to a component that utilizes the multilayer perceptron (MLP) architecture in deep learning to process input data and identify internal structural features. It is primarily used to analyze the internal structural connections of the hydrogen energy storage device and generate an energy transfer path topology map. This module simulates how energy flows within the system by learning the complex relationships between data from different sensors. The gas diffusion physical constraint module is a module designed based on physics principles to simulate and predict the diffusion behavior of a gas (hydrogen in this case) under specific environmental conditions. This module combines real-time environmental parameters (such as temperature, wind speed, etc.) to construct a three-dimensional spatial grid coordinate system and applies the mass conservation equation and momentum transfer theorem to calculate the gas diffusion path. The probability density clustering module is a data analysis tool specifically designed to identify dense regions or clusters in high-dimensional datasets. In hydrogen energy storage applications, it is used to analyze and merge leakage situation maps, identify areas with the highest leakage risk, and generate leakage probability cloud maps. Furthermore, it can construct energy loss vectors based on the energy flow field feature matrix, helping to quantify the degree of energy loss.

[0100] The step of inputting the fused feature tensor into the spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage equipment, and outputting a leakage probability cloud map and an energy loss vector, includes:

[0101] Step 201: Analyze the internal structural connection relationship of the hydrogen energy storage device through the multi-layer sensing constraint module to generate an energy transfer path topology map.

[0102] In step 201, the energy transfer path topology diagram refers to a network diagram that represents the energy transfer path within the device, where nodes represent device components and edges represent energy transfer relationships.

[0103] In this embodiment, the connection relationships within the internal structure of the hydrogen energy storage device are analyzed using a multi-layer perceptual constraint module. A graph neural network is then used to model the energy transfer relationships between the device components, generating an energy transfer path topology graph. The node weights of this topology graph are determined by the energy transfer efficiency of the device components, while the edge weights are determined by the connection strength between the components.

[0104] Step 202: Based on the energy transfer path topology, the energy flow field feature matrix is ​​tracked by the dynamic recursive operator in the implicit state transition model. The dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform.

[0105] In step 202, the dynamic recursive operator is an operator used to perform state transition calculations on the energy transfer path topology graph, which can adaptively adjust the step size.

[0106] In this embodiment, the temperature gradient is decomposed into thermal conduction and thermal stress components by thermoelastic coupling equation, and a mechanical vibration influence factor matrix is ​​generated by combining the frequency domain energy distribution of the vibration spectrum. Finally, a temperature vibration coupling coefficient matrix is ​​generated to adjust the state transition step size.

[0107] Step 203: Input the energy flow field feature matrix into the gas diffusion physics constraint module, and construct a three-dimensional spatial grid coordinate system in combination with real-time environmental parameters.

[0108] In step 203, the three-dimensional spatial grid coordinate system refers to dividing the space around the device into a three-dimensional grid for simulating gas diffusion.

[0109] In this embodiment, the energy flow field feature matrix is ​​input into the gas diffusion physical constraint module, and a three-dimensional spatial grid coordinate system is constructed by combining real-time environmental parameters (such as wind speed, air pressure, and temperature). This coordinate system divides the space around the device into several grid cells, each containing parameters such as gas concentration, temperature, and pressure.

[0110] Step 204: Simulate the diffusion path of the leaked gas in the three-dimensional spatial grid coordinate system using the mass conservation equation and momentum transfer theorem to generate the probability distribution of the diffusion path of the leaked gas.

[0111] In step 204, the diffusion path probability distribution is used to represent the distribution of the diffusion probability of gas in different grid cells.

[0112] In this embodiment, the mass change and momentum transfer of the gas in each grid cell are calculated to generate the probability distribution of the diffusion path of the leaked gas.

[0113] Step 205: Generate a fusion leakage situation map based on the diffusion path probability distribution.

[0114] In step 205, the fused leak situation map is the result of a comprehensive analysis of all collected information, showing the possible location and severity of the leak.

[0115] In this embodiment, a fusion leakage situation map is generated based on the diffusion path probability distribution. This map combines gas diffusion path and energy flow field characteristics, and can intuitively reflect the distribution and diffusion trend of leakage risk.

[0116] Step 206: Based on the fused leakage situation map, generate a leakage probability cloud map through the probability density clustering module, and construct an energy loss vector according to the energy flow field feature matrix.

[0117] In this embodiment, a Gaussian mixture model is used to perform cluster analysis on the leakage situation map to generate a leakage probability cloud map; at the same time, an energy loss vector is constructed based on the energy flow field characteristic matrix to quantify the energy loss caused by leakage.

[0118] Here is a specific example:

[0119] In a hydrogen energy storage device, a multi-layer sensing constraint module analyzes the internal structure of the device and generates an energy transfer path topology map; an implicit state transition model tracks the energy flow field feature matrix and discovers an abnormal coupling between the temperature gradient and vibration spectrum in a certain region; a gas diffusion physical constraint module constructs a three-dimensional spatial grid coordinate system to simulate the diffusion path of leaked gas and generates a diffusion path probability distribution; based on the diffusion path probability distribution, a fused leakage situation map is generated, showing that the leaked gas diffuses towards the southeast of the device; a probability density clustering module generates a leakage probability cloud map, finding that the leakage probability in the southeast region is as high as 80%, and simultaneously constructs an energy loss vector, showing an increase in energy loss in this region. The system triggers a secondary early warning signal, generates a leak source thermal map, and dynamically adjusts the valve openings of adjacent devices to balance the pressure distribution.

[0120] This scheme achieves refined prediction and dynamic control of leaks in hydrogen energy storage devices through the collaborative work of a multi-layered sensing constraint module, an implicit state transition model, a gas diffusion physical constraint module, and a probability density clustering module. Specifically, the generated energy transfer path topology map and energy flow field feature matrix provide the data foundation for leak prediction. The leak probability cloud map and energy loss vector generated by the gas diffusion physical constraint module and the probability density clustering module improve the accuracy of leak location and energy loss analysis. Finally, the self-regulating system dynamically optimizes the device's operating parameters, forming a closed-loop control link, effectively improving the safety and operating efficiency of the hydrogen energy storage device.

[0121] To further improve the accuracy and dynamic adjustment capability of energy flow field feature tracking in hydrogen energy storage devices, in some embodiments, step 202: based on the energy transfer path topology map, the energy flow field feature matrix is ​​tracked through a dynamic recursive operator in the implicit state transition model. The dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform, including:

[0122] Step 301: Decompose the temperature gradient distribution field into heat conduction components and thermal stress components by using a preset thermoelastic coupling equation, and generate a mechanical vibration influence factor matrix based on the frequency domain energy distribution of the vibration spectrum waveform.

[0123] In step 301, the preset thermoelastic coupling equation is a mathematical model used to describe the influence of temperature change (heat) on the internal stress distribution (elasticity) of the material, and the interaction between the two. The heat conduction component refers to the temperature change caused by heat conduction in the temperature gradient distribution field. The thermal stress component refers to the temperature change caused by thermal stress in the temperature gradient distribution field. The mechanical vibration influence factor matrix is ​​a matrix representing the influence of the vibration spectrum waveform on the mechanical vibration of the equipment.

[0124] In this embodiment, the temperature gradient distribution field is decomposed into heat conduction components and thermal stress components by a preset thermoelastic coupling equation, which respectively represent the temperature changes caused by heat conduction and thermal stress. At the same time, based on the frequency domain energy distribution of the vibration spectrum waveform, the energy values ​​of the main frequency bands are extracted using fast Fourier transform to generate a mechanical vibration influence factor matrix, which is used to quantify the impact of vibration on the equipment.

[0125] Step 302: Perform a dual-channel convolution fusion operation on the heat conduction component and the mechanical vibration influence factor matrix to generate a temperature vibration coupling coefficient matrix.

[0126] In step 302, the dual-channel convolution fusion operation is a technique that fuses two feature matrices through convolution. The temperature-vibration coupling coefficient matrix is ​​a matrix representing the coupling effect between the temperature gradient and the vibration spectrum, and its dimension is consistent with the number of nodes in the energy transfer path topology graph.

[0127] In this embodiment, a dual-channel convolution fusion operation is performed on the thermal conduction component and the mechanical vibration influence factor matrix. A convolutional neural network is used to perform convolution processing on the two feature matrices respectively to extract their spatial features. Then, a temperature-vibration coupling coefficient matrix is ​​generated through feature splicing and a fully connected layer. Each element of this matrix represents the temperature-vibration coupling strength of the corresponding node.

[0128] Step 303: Perform a state transition calculation with physical constraints on the energy transfer path topology graph, and iteratively update the energy flow field feature matrix. In each iteration, the dynamic recursive operator dynamically adjusts the state transition step size according to the spectral radius of the temperature vibration coupling coefficient matrix.

[0129] In step 303, the state transition calculation with physical constraints refers to introducing physical constraints (such as energy conservation and momentum conservation) into the state transition calculation to ensure the rationality of the calculation results. The spectral radius refers to the maximum absolute value of the matrix eigenvalues, used to measure the stability of the matrix. The state transition step size refers to the step size of each iteration in the state transition calculation, affecting computational efficiency and accuracy.

[0130] In this embodiment, a physically constrained state transition calculation is performed on the energy transfer path topology to iteratively update the energy flow field characteristic matrix. Specifically, in each iteration, the dynamic recursive operator dynamically adjusts the state transition step size based on the spectral radius of the temperature vibration coupling coefficient matrix: if the spectral radius is large, the step size is reduced to improve stability; if the spectral radius is small, the step size is increased to improve computational efficiency. Finally, an updated energy flow field characteristic matrix is ​​generated.

[0131] Here is a specific example:

[0132] In a hydrogen energy storage station application scenario, the monitored temperature gradient distribution field is first divided into heat conduction and thermal stress components using a thermoelastic coupling equation, and a mechanical vibration influence factor matrix is ​​extracted from vibration sensor data. Next, a dual-channel convolutional fusion algorithm is used to process this data to generate a temperature-vibration coupling coefficient matrix. Finally, a physically constrained state transition calculation is applied to the energy transfer path topology map, dynamically adjusting the state transition step size to accurately track changes in the energy flow field, thus improving the accuracy and timeliness of leakage prediction.

[0133] This scheme achieves refined coupling analysis of temperature gradient and vibration spectrum through thermoelastic coupling equation and dual-channel convolution fusion operation, generating a temperature-vibration coupling coefficient matrix. Combined with physically constrained state transition calculation and dynamic recursive operator, it realizes efficient iterative update of energy flow field feature matrix, improves the accuracy of energy flow field feature tracking and dynamic adjustment capability, and provides a reliable data foundation for subsequent leakage prediction and energy loss analysis.

[0134] To further improve the accuracy and rationality of state transition calculations, in some embodiments, step 303, before performing physically constrained state transition calculations on the energy transfer path topology graph, further includes:

[0135] Step 401: Perform frequency domain energy spectrum analysis on the pressure fluctuation sequence to obtain the analysis results.

[0136] In step 401, the analysis result refers to the energy distribution of the pressure fluctuation sequence in the frequency domain.

[0137] In this embodiment, frequency domain energy spectrum analysis is performed on the pressure fluctuation sequence. Fourier transform is used to convert the time domain pressure signal into a frequency domain energy spectrum to obtain the energy distribution at different frequencies, providing a data basis for the subsequent extraction of the energy mutation index.

[0138] Step 402: Extract the energy mutation index in different frequency bands from the analysis results through a bandpass filter window, and construct the pressure frequency domain mutation feature vector.

[0139] In step 402, the bandpass filter window is a signal processing term referring to a filter capable of selecting signal components within a specific frequency range (i.e., a frequency band). The energy mutation index represents the intensity of an energy mutation within a certain frequency band. The pressure frequency domain mutation eigenvector is a vector composed of energy mutation indices from different frequency bands.

[0140] In this embodiment, the energy mutation index in different frequency bands is extracted from the frequency domain energy spectrum through a bandpass filter window, the frequency domain energy spectrum is segmented using a sliding window technique, the energy mutation index of each frequency band is calculated, and finally the pressure frequency domain mutation feature vector is constructed.

[0141] Step 403: Calculate the spatial gradient direction weight coefficient based on the preset leakage diffusion direction constraint conditions.

[0142] In step 403, the leakage diffusion direction constraint refers to the directional restriction on how the leakage diffuses, set based on the physical model and historical data. The spatial gradient direction weighting coefficient is a weighting coefficient representing the influence of the gas diffusion direction on pressure fluctuations.

[0143] In this embodiment, the gradient direction weighting coefficient for each spatial location is calculated based on the physical laws of gas diffusion (such as diffusion from high pressure to low pressure) and the internal pressure distribution of the device.

[0144] Step 404: Input the pressure frequency domain abrupt change feature vector and the spatial gradient direction weight coefficient into the dual-channel attention gating module to generate the initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module.

[0145] In step 404, the dual-channel attention gating module refers to a module that combines an attention mechanism to establish the correlation between pressure frequency domain features and spatial gradient features. The initial value of the time-varying gain factor refers to the initial value representing the correlation strength between pressure frequency domain features and spatial gradient features.

[0146] In this embodiment, the pressure frequency domain abrupt change feature vector and the spatial gradient direction weight coefficient are input into the dual-channel attention gating module, and the correlation mapping between the two is established through a nonlinear mapping function to generate the initial value of the time-varying gain factor.

[0147] Step 405: Use the leakage risk index threshold surface to dynamically normalize the initial value of the time-varying gain factor to generate the normalized time-varying gain factor.

[0148] In step 405, the leakage risk index threshold surface is the threshold surface used for dynamically normalizing the time-varying gain factor. The normalized time-varying gain factor refers to the time-varying gain factor after dynamic normalization, and its value range is [0,1].

[0149] In this embodiment of the application, the leakage risk index threshold surface is used to dynamically normalize the initial value of the time-varying gain factor, mapping the initial value of the time-varying gain factor to the range of [0,1], and generating the normalized time-varying gain factor.

[0150] Step 406: Perform tensor multiplication on the normalized time-varying gain factor and the temperature vibration coupling coefficient matrix to generate a dynamic weight parameter matrix.

[0151] In step 406, the dynamic weight parameter matrix refers to the weight matrix obtained by combining the time-varying gain factor and the temperature vibration coupling coefficient matrix.

[0152] In this embodiment, the normalized time-varying gain factor and the temperature vibration coupling coefficient matrix are multiplied by a tensor to generate a dynamic weight parameter matrix. This matrix is ​​used to adjust the weight allocation in the state transition calculation.

[0153] Step 407: The dynamic weight parameter matrix is ​​embedded as a physical constraint into the initial state transition equation to obtain the state transition equation with physical constraints.

[0154] In step 407, the physically constrained state transition equation is a modified traditional state transition equation that takes into account physical reality and incorporates all the information obtained from the previous steps.

[0155] In this embodiment, the dynamic weight parameter matrix is ​​embedded as a physical constraint into the initial state transition equation, resulting in a state transition equation with physical constraints. This equation considers the coupling effect of pressure fluctuations, temperature gradients, and vibration spectra in the state transition calculation, improving the accuracy and rationality of the calculation.

[0156] Here is a specific example:

[0157] In a hydrogen energy storage station application scenario, the pressure fluctuation sequence is first analyzed using frequency domain energy spectrum analysis to extract energy mutation indices in different frequency bands and construct pressure frequency domain mutation feature vectors. Next, spatial gradient direction weighting coefficients are calculated based on leakage diffusion direction constraints, and initial values ​​of time-varying gain factors are generated using a dual-channel attention gating module. Then, these initial values ​​are normalized using a leakage risk index threshold surface, and finally combined with the temperature vibration coupling coefficient matrix to generate a dynamic weighting parameter matrix, which is applied to the state transition equation. This achieves high-precision tracking of energy flow patterns and significantly improves the effectiveness of the leakage early warning system.

[0158] This scheme achieves refined coupled analysis of pressure fluctuations, temperature gradients, and vibration spectra through frequency domain energy spectrum analysis, bandpass filtering, spatial gradient direction weight calculation, and dual-channel attention gating modules. By generating a dynamic weight parameter matrix through dynamic normalization and tensor multiplication operations and embedding it into the state transition equation, the accuracy and rationality of the state transition calculation are improved, providing reliable technical support for leakage prediction and energy loss analysis of hydrogen energy storage equipment.

[0159] To further improve the correlation mapping accuracy between pressure frequency domain features and spatial gradient features, in some embodiments, step 404: inputting the pressure frequency domain abrupt change feature vector and the spatial gradient direction weight coefficient into a dual-channel attention gating module to generate an initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module, includes:

[0160] Step 501: Perform a multi-scale decomposition operation on the pressure frequency domain abrupt change feature vector to obtain multi-scale decomposition information.

[0161] In step 501, the multi-scale decomposition information includes low-frequency, mid-frequency, and high-frequency scales. The low-frequency scale captures the global trend and slow-changing features in the pressure fluctuation sequence, the mid-frequency scale reflects the energy distribution and fluctuation features in the mid-frequency range, and the high-frequency scale extracts rapidly changing details and local abrupt changes.

[0162] In this embodiment of the application, in order to perform multi-scale decomposition on the pressure frequency domain mutation feature vector, empirical mode decomposition or wavelet transform is used to decompose the pressure frequency domain mutation feature vector into sub-signals of multiple scales to obtain multi-scale decomposition information, which provides a data basis for the subsequent extraction of energy mutation intensity coefficients.

[0163] Step 502: Perform wavelet packet transform on the multi-scale decomposition information to extract the energy mutation intensity coefficients of different frequency bands.

[0164] In step 502, wavelet packet transform refers to a more refined frequency domain analysis method than wavelet transform, capable of extracting more detailed frequency band features. The energy jump intensity coefficient represents the intensity of energy jumps within a certain frequency band.

[0165] In this embodiment, wavelet packet transform is used to perform frequency domain analysis on sub-signals at each scale, and the energy mutation intensity coefficient of each frequency band is calculated.

[0166] Step 503: Construct the pressure frequency domain energy distribution tensor based on the energy mutation intensity coefficients of different frequency bands.

[0167] In step 503, the pressure frequency domain energy distribution tensor is a high-dimensional tensor composed of energy abrupt change intensity coefficients in different frequency bands, representing the distribution of pressure frequency domain energy.

[0168] In this embodiment, a pressure frequency domain energy distribution tensor is constructed based on the energy mutation intensity coefficients of different frequency bands. Each dimension of this tensor corresponds to the energy mutation intensity coefficient of a frequency band, which is used to represent the multi-scale distribution characteristics of pressure frequency domain energy.

[0169] Step 504: Perform directional enhancement processing on the spatial gradient direction weight coefficients to obtain the processed spatial gradient direction weight coefficients.

[0170] In step 504, the processed spatial gradient direction weight coefficients refer to the weight coefficients after directional enhancement processing.

[0171] In this embodiment, the spatial gradient direction weight coefficients are subjected to directional enhancement processing. A Gaussian filtering algorithm is used to smooth the weight coefficients, while the weight values ​​of the main diffusion directions are enhanced to obtain the processed spatial gradient direction weight coefficients.

[0172] Step 505: Based on the processed spatial gradient direction weight coefficients, an anisotropic diffusion filtering algorithm is used to generate a gradient direction consistency matrix.

[0173] In step 505, the gradient direction consistency matrix represents the matrix of spatial gradient direction consistency.

[0174] In this embodiment, an anisotropic diffusion filtering algorithm is used to generate a gradient direction consistency matrix based on the processed spatial gradient direction weight coefficients. This algorithm adjusts the filtering intensity according to the gradient direction to generate the consistency matrix, which is used to quantify the consistency of the gradient direction.

[0175] Step 506: Calculate the direction weight correction factor based on the gradient direction consistency matrix and the preset leakage diffusion direction constraint conditions.

[0176] In step 506, the orientation weight correction factor is used to correct the spatial gradient orientation weights to improve the rationality of weight allocation.

[0177] In this embodiment, a direction weight correction factor is calculated based on the gradient direction consistency matrix and preset leakage diffusion direction constraints. Specifically, the weight allocation is adjusted according to the consistency matrix and leakage diffusion direction constraints to generate the direction weight correction factor.

[0178] Step 507: Input the pressure frequency domain energy distribution tensor and the direction weight correction factor into the dual-channel attention gating module to establish a nonlinear correlation mapping between the pressure frequency domain and the spatial gradient through the cross attention module in the dual-channel attention gating module, and generate an initial attention weight matrix.

[0179] In step 507, the initial attention weight matrix represents the initial matrix of the correlation strength between the pressure frequency domain and the spatial gradient features.

[0180] In this embodiment, the pressure frequency domain energy distribution tensor and the direction weight correction factor are input into a dual-channel attention gating module. A nonlinear correlation mapping between the pressure frequency domain and the spatial gradient is established through a cross-attention module, generating an initial attention weight matrix. This matrix represents the correlation strength between the two.

[0181] Step 508: Perform a nonlinear activation operation on the initial attention weight matrix to generate initial values ​​for the time-varying gain factor.

[0182] In this embodiment, a nonlinear activation operation is performed on the initial attention weight matrix to conduct a nonlinear transformation of the matrix, generating an initial value for the time-varying gain factor. This value is used for subsequent dynamic normalization processing.

[0183] Here is a specific example:

[0184] In a hydrogen energy storage device, a multi-scale decomposition operation is performed on the pressure frequency domain abrupt change feature vector to obtain multi-scale decomposition information; energy abrupt change intensity coefficients of different frequency bands are extracted by wavelet packet transform to construct a pressure frequency domain energy distribution tensor; the spatial gradient direction weight coefficients are subjected to directionality enhancement processing to generate a gradient direction consistency matrix; the direction weight correction factor is calculated in combination with leakage diffusion direction constraints; the pressure frequency domain energy distribution tensor and the direction weight correction factor are input into a dual-channel attention gating module to generate an initial attention weight matrix; and an initial value of the time-varying gain factor is generated through a nonlinear activation operation for subsequent state transition calculations.

[0185] This scheme achieves a refined correlation mapping between pressure frequency domain features and spatial gradient features through multi-scale decomposition, wavelet packet transform, directional enhancement processing, and cross-attention module. The generated pressure frequency domain energy distribution tensor and directional weight correction factor improve the accuracy of feature expression. The initial value of the finally generated time-varying gain factor provides a high-precision weight allocation basis for subsequent state transition calculations, further improving the accuracy and reliability of leakage prediction for hydrogen energy storage equipment.

[0186] To further improve the fusion effect and feature representation capability of the multidimensional feature matrix, in some embodiments, step 102: inputting the multidimensional feature matrix into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross-attention module in the bidirectional interactive network, and using a gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor, including:

[0187] Step 601: Perform time-frequency joint analysis on the pressure fluctuation sequence to generate a pressure fluctuation time-frequency feature matrix.

[0188] In step 601, time-frequency joint analysis refers to simultaneously analyzing the characteristics of the signal in both the time and frequency domains to capture the dynamic changes and frequency distribution of the signal. The pressure fluctuation time-frequency feature matrix is ​​a high-dimensional matrix representing the joint characteristics of the pressure fluctuation sequence in both the time and frequency domains.

[0189] In this embodiment, a joint time-frequency analysis is performed on the pressure fluctuation sequence to transform it from the time domain to the time-frequency domain, generating a pressure fluctuation time-frequency feature matrix. This matrix contains information in three dimensions: time, frequency, and energy, and is used to comprehensively describe the dynamic characteristics of the pressure fluctuations.

[0190] Step 602: Perform spatial multi-resolution analysis on the gas concentration spectrum to generate a gas concentration multi-scale feature map.

[0191] In step 602, spatial multi-resolution analysis refers to decomposing the gas concentration spectrum at different spatial scales to capture multi-level concentration distribution characteristics. A gas concentration multi-scale feature map is a map representing the characteristic distribution of the gas concentration spectrum at different spatial scales. The multi-scale features map includes large-scale, medium-scale, and small-scale features. Large-scale features capture the distribution trend and global characteristics of gas concentration within the overall spatial range of the equipment; medium-scale features reflect the variation characteristics and transition regions of gas concentration within local areas; and small-scale features extract detailed information and local abrupt changes in gas concentration within a fine spatial range.

[0192] In this embodiment, spatial multi-resolution analysis is performed on the gas concentration spectrum. The Gaussian pyramid algorithm is used to decompose the gas concentration spectrum into feature maps of multiple spatial scales, generating a multi-scale feature map of gas concentration. This map contains concentration distribution information at different resolutions, used to describe the spatial variation characteristics of gas concentration.

[0193] Step 603: Input the pressure fluctuation time-frequency feature matrix and the gas concentration multi-scale feature map into the cross-attention module, so as to establish the correlation mapping between the pressure frequency domain features and the gas concentration spatial features through the multi-head attention module, and generate the pressure-concentration interaction feature matrix.

[0194] In step 603, the pressure-concentration interaction feature matrix represents a high-dimensional matrix that maps the pressure frequency domain features to the gas concentration spatial features.

[0195] In this embodiment, the pressure fluctuation time-frequency feature matrix and the gas concentration multi-scale feature map are input into the cross-attention module, and the correlation mapping between the pressure frequency domain features and the gas concentration spatial features is established through the multi-head attention module. Specifically, the correlation weight between the two is calculated using the multi-head attention mechanism to generate the pressure-concentration interaction feature matrix.

[0196] Step 604: Generate a temperature gradient propagation path map based on the temperature gradient distribution field.

[0197] In step 604, the temperature gradient propagation path diagram represents the path of the temperature gradient propagation inside the device.

[0198] In this embodiment, a temperature gradient propagation path map is generated based on the temperature gradient distribution field. The heat conduction equation is used to simulate the propagation path of the temperature gradient inside the device, thus generating the temperature gradient propagation path map. This map is used to describe the spatial propagation characteristics of the temperature gradient.

[0199] Step 605: Extract the intrinsic mode function components of the vibration spectrum waveform using the empirical mode decomposition algorithm, and generate a vibration spectrum energy feature vector based on the intrinsic mode function components.

[0200] In step 605, the intrinsic mode function components represent the components of different frequencies in the signal. The vibration spectrum energy eigenvector represents the eigenvector of the vibration spectrum waveform energy distribution.

[0201] In this embodiment, the intrinsic mode function components of the vibration spectrum waveform are extracted by the empirical mode decomposition algorithm. The vibration spectrum waveform is decomposed into multiple intrinsic mode function components by empirical mode decomposition, and a vibration spectrum energy feature vector is generated based on these components.

[0202] Step 606: Generate a temperature vibration interaction feature matrix based on the temperature gradient propagation path diagram and the vibration spectrum energy feature vector.

[0203] In step 606, the temperature-vibration interaction feature matrix represents a high-dimensional matrix that maps the temperature gradient to the vibration spectrum features.

[0204] In this embodiment, a temperature-vibration interaction feature matrix is ​​generated based on the temperature gradient propagation path map and the vibration spectrum energy feature vector. Specifically, a convolutional neural network is used to fuse the features of the two to generate the temperature-vibration interaction feature matrix.

[0205] Step 607: Perform a tensor concatenation operation on the pressure-concentration interaction feature matrix and the temperature-vibration interaction feature matrix to generate a fused feature tensor.

[0206] In this embodiment, the pressure-concentration interaction feature matrix and the temperature-vibration interaction feature matrix are concatenated using a tensor operation to generate a fused feature tensor. This tensor contains multi-source feature information of pressure, gas concentration, temperature, and vibration, which is used for subsequent leakage prediction and energy loss analysis.

[0207] Here is a specific example:

[0208] In a hydrogen energy storage device, time-frequency joint analysis is performed on the pressure fluctuation sequence to generate a pressure fluctuation time-frequency feature matrix; spatial multi-resolution analysis is performed on the gas concentration spectrum to generate a gas concentration multi-scale feature map; both are input into a cross-attention module to generate a pressure-concentration interaction feature matrix; a temperature gradient propagation path map is generated based on the temperature gradient distribution field; intrinsic mode function components of the vibration spectrum waveform are extracted using an empirical mode decomposition algorithm to generate a vibration spectrum energy feature vector; a temperature-vibration interaction feature matrix is ​​generated based on the temperature gradient propagation path map and the vibration spectrum energy feature vector; finally, the pressure-concentration interaction feature matrix and the temperature-vibration interaction feature matrix are concatenated to generate a fused feature tensor, providing high-dimensional feature input for subsequent leakage prediction.

[0209] This scheme achieves deep fusion of multi-source features such as pressure, gas concentration, temperature, and vibration through joint time-frequency analysis, spatial multi-resolution analysis, cross-attention modules, and empirical mode decomposition algorithms. The generated fused feature tensor improves the comprehensiveness and accuracy of feature representation, providing high-dimensional and high-precision feature inputs for leakage prediction and energy loss analysis of hydrogen energy storage equipment, and further enhancing the accuracy and reliability of the prediction model.

[0210] To further improve the accuracy and dynamic control capability of hydrogen energy storage equipment leakage early warning, in some embodiments, step 104: determining the entropy value mutation region based on the energy loss vector, when the high-density region in the leakage probability cloud map and the entropy value mutation region spatially overlap within a continuous detection period, triggering a multi-level early warning signal and generating a leakage source thermal map and pressure balance parameter set, including:

[0211] Step 701: Perform a multi-dimensional decomposition operation on the energy loss vector to generate an energy loss feature matrix.

[0212] In step 701, the energy loss feature matrix represents the characteristic distribution matrix of energy loss in multiple dimensions. The multiple dimensions in the multi-dimensional decomposition operation include time, space, and energy dimensions. The time dimension analyzes the trend of energy loss over time, the spatial dimension describes the distribution characteristics of energy loss at different spatial locations of the device, and the energy dimension reflects the distribution of energy loss at different energy levels. Through multi-dimensional decomposition, various characteristics of energy loss can be comprehensively captured, providing multi-dimensional feature support for subsequent identification of entropy abrupt change regions.

[0213] In this embodiment, a multi-dimensional decomposition operation is performed on the energy loss vector. Independent component analysis is used to decompose the energy loss vector into features of multiple dimensions, generating an energy loss feature matrix. This matrix is ​​used to describe the distribution characteristics of energy loss in different dimensions.

[0214] Step 702: Identify abnormal energy entropy values ​​based on the energy loss feature matrix to generate entropy value mutation regions.

[0215] In step 702, the abnormal energy entropy value represents the region in the energy loss feature matrix where the entropy value is abnormal.

[0216] In this embodiment, abnormal energy entropy values ​​are identified based on the energy loss feature matrix, and an entropy calculation algorithm is used to analyze the energy loss feature matrix to identify regions with abnormal entropy values ​​and generate regions with abrupt entropy changes.

[0217] Step 703: Perform spatial clustering analysis on the leakage probability cloud map to generate high-density regions.

[0218] In this embodiment, spatial clustering analysis is performed on the leakage probability cloud map to identify high-density regions. These regions represent areas with a higher leakage probability than other regions.

[0219] Step 704: When the high-density region and the entropy change region spatially overlap within at least two consecutive detection cycles, a multi-level early warning signal is triggered.

[0220] In this embodiment, when a high-density region and an entropy abrupt change region spatially overlap within at least two consecutive detection cycles, a multi-level early warning signal is triggered. Specifically, the risk level is determined based on the area of ​​the overlapping region and the leakage probability value, and an early warning signal of the corresponding level is triggered.

[0221] Step 705: Generate a leakage source thermal map corresponding to the high-density area. Based on the leakage source thermal map and the preset pressure balance equation, calculate the pressure compensation parameters of multiple adjacent hydrogen energy storage devices to generate a pressure balance parameter set.

[0222] In this embodiment, a thermal map of the leakage source corresponding to the high-density area is generated. Based on the thermal map of the leakage source and a preset pressure balance equation, pressure compensation parameters for multiple adjacent hydrogen energy storage devices are calculated to generate a pressure balance parameter set. This parameter set is used to dynamically adjust the valve opening of adjacent devices to balance the pressure distribution.

[0223] Here is a specific example:

[0224] In a hydrogen energy storage device, a multi-dimensional decomposition operation is performed on the energy loss vector to generate an energy loss feature matrix; abnormal energy entropy values ​​are identified based on the energy loss feature matrix to generate entropy value mutation regions; spatial clustering analysis is performed on the leakage probability cloud map to generate high-density regions; when the high-density regions overlap with the entropy value mutation regions within two consecutive detection cycles, a secondary early warning signal is triggered, and a leakage source heat map is generated; based on the leakage source heat map and the pressure balance equation, the pressure compensation parameters of adjacent equipment are calculated to generate a pressure balance parameter set for dynamically adjusting valve opening.

[0225] This solution achieves accurate early warning and dynamic control of leakage risks in hydrogen energy storage equipment through multi-dimensional decomposition, entropy analysis, spatial clustering, and multi-level early warning mechanisms. The generated leakage source thermal map and pressure balance parameter set improve the accuracy of leakage location and pressure balance, providing reliable technical support for the safe operation of the equipment.

[0226] Figure 2 A schematic diagram of a deep learning-based hydrogen energy storage device leakage fault early warning system provided in this application embodiment is shown below. Figure 2 As shown, the system includes:

[0227] The collection module 21 is used to collect multi-source heterogeneous sensor data streams from hydrogen energy storage devices in real time to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra.

[0228] Input module 22 is used to input the multidimensional feature matrix into the bidirectional interactive network, so as to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network, and to use the gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor.

[0229] Output module 23 is used to input the fused feature tensor into the spatiotemporal inference model, so as to use the spatiotemporal inference model to predict the leakage and energy loss of hydrogen energy storage equipment, and output the leakage probability cloud map and energy loss vector.

[0230] Trigger module 24 is used to determine the entropy value mutation region based on the energy loss vector. When the high-density region in the leakage probability cloud map and the entropy value mutation region spatially overlap within a continuous detection period, it triggers a multi-level early warning signal and generates a leakage source thermal map and pressure balance parameter set.

[0231] The activation module 25 is used to activate the self-regulating system according to the leakage source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

[0232] Figure 2 The aforementioned deep learning-based hydrogen energy storage device leakage fault early warning system can perform... Figure 1 The implementation principle and technical effects of the deep learning-based hydrogen energy storage device leakage fault early warning method described in the illustrated embodiment will not be repeated here. The specific operation methods of each module and unit in the deep learning-based hydrogen energy storage device leakage fault early warning system described in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0233] In one possible design, Figure 2 The deep learning-based hydrogen energy storage device leakage fault early warning system of the embodiment shown can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0234] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0235] The processing component 32 is described above Figure 1 The embodiment describes a method for early warning of leakage faults in hydrogen energy storage equipment based on deep learning.

[0236] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0237] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0238] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0239] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0240] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0241] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0242] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is a method for early warning of leakage faults in hydrogen energy storage equipment based on deep learning.

[0243] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0244] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0245] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0246] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for early warning of leakage faults in hydrogen energy storage equipment based on deep learning, characterized in that, include: Multi-source heterogeneous sensor data streams from hydrogen energy storage devices are collected in real time to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra. The multidimensional feature matrix is ​​input into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network. The gradient field reconstruction algorithm is used to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor. The fused feature tensor is input into the spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage equipment, and outputs a leakage probability cloud map and an energy loss vector. Based on the energy loss vector, the entropy value mutation region is determined. When the high-density region in the leakage probability cloud map and the entropy value mutation region spatially overlap within a continuous detection period, a multi-level early warning signal is triggered and a leakage source thermal map and pressure balance parameter set are generated. The entropy value mutation region refers to the region in the energy loss vector where the entropy value changes, indicating abnormal energy fluctuations. The self-regulating system is activated based on the leak source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

2. The method according to claim 1, characterized in that, The spatiotemporal reasoning model includes a multi-layer perception constraint module, an implicit state transition model, a gas diffusion physical constraint module, and a probability density clustering module connected in sequence. The step of inputting the fused feature tensor into the spatiotemporal inference model to predict leakage and energy loss of hydrogen energy storage equipment, and outputting a leakage probability cloud map and an energy loss vector, includes: The internal structural connections of the hydrogen energy storage device are analyzed by a multi-layer perceptual constraint module to generate an energy transfer path topology map. The multi-layer perceptual constraint module refers to a component that uses the multi-layer perceptron architecture in deep learning to process input data and identify internal structural features. Based on the energy transfer path topology, the energy flow field feature matrix is ​​tracked by a dynamic recursive operator in the implicit state transition model. The dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform. The implicit state transition model is a model used to perform state transition calculations on the energy transfer path topology to track the energy flow field features. The energy flow field feature matrix is ​​a feature matrix used to characterize the energy flow state inside the device, obtained by tracking the energy transfer path topology through the implicit state transition model. The energy flow field feature matrix is ​​input into the gas diffusion physics constraint module, and a three-dimensional spatial grid coordinate system is constructed by combining real-time environmental parameters. The gas diffusion physics constraint module is a module designed based on physical principles to simulate and predict the diffusion behavior of gas under specific environmental conditions. The diffusion path of leaked gas in a three-dimensional spatial grid coordinate system is simulated by the mass conservation equation and the momentum transfer theorem, and the probability distribution of the diffusion path of leaked gas is generated. A fusion leakage situation map is generated based on the probability distribution of the diffusion path. Based on the fused leakage situation map, a leakage probability cloud map is generated by the probability density clustering module, and an energy loss vector is constructed according to the energy flow field feature matrix. The probability density clustering module is a data analysis tool used to identify dense regions or clusters in high-dimensional data.

3. The method according to claim 2, characterized in that, Based on the energy transfer path topology, the energy flow field feature matrix is ​​tracked by a dynamic recursive operator in the implicit state transition model. The dynamic recursive operator adaptively adjusts the state transition step size according to the coupling effect between the temperature gradient distribution field and the vibration spectrum waveform, including: The temperature gradient distribution field is decomposed into heat conduction and thermal stress components by a preset thermoelastic coupling equation, and a mechanical vibration influence factor matrix is ​​generated based on the frequency domain energy distribution of the vibration spectrum waveform. A dual-channel convolution fusion operation is performed on the heat conduction component and the mechanical vibration influence factor matrix to generate a temperature vibration coupling coefficient matrix. The dimension of the coupling coefficient matrix is ​​consistent with the number of nodes in the energy transfer path topology graph. A physically constrained state transition calculation is performed on the energy transfer path topology, and the energy flow field feature matrix is ​​iteratively updated. In each iteration, the dynamic recursive operator dynamically adjusts the state transition step size according to the spectral radius of the temperature vibration coupling coefficient matrix.

4. The method according to claim 3, characterized in that, Before performing physically constrained state transition calculations on the energy transfer path topology, the method further includes: Frequency domain energy spectrum analysis was performed on the pressure fluctuation sequence to obtain the analysis results; Energy mutation indexes in different frequency bands are extracted from the analysis results using a bandpass filter window to construct a pressure frequency domain mutation feature vector. And calculate the spatial gradient direction weighting coefficient based on the preset leakage diffusion direction constraint conditions; The pressure frequency domain abrupt change feature vector and the spatial gradient direction weight coefficient are input into the dual-channel attention gating module to generate the initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module; The initial value of the time-varying gain factor is dynamically normalized using the leakage risk index threshold surface to generate a normalized time-varying gain factor, wherein the leakage risk index threshold surface is a threshold surface used to dynamically normalize the time-varying gain factor. The normalized time-varying gain factor is multiplied by the temperature vibration coupling coefficient matrix to generate a dynamic weight parameter matrix. By embedding the dynamic weight parameter matrix as a physical constraint into the initial state transition equation, a state transition equation with physical constraints is obtained.

5. The method according to claim 4, characterized in that, The step of inputting the pressure frequency domain abrupt change feature vector and the spatial gradient direction weighting coefficient into the dual-channel attention gating module to generate the initial value of the time-varying gain factor through the nonlinear mapping function in the dual-channel attention gating module includes: Perform a multi-scale decomposition operation on the pressure frequency domain abrupt change feature vector to obtain multi-scale decomposition information; Wavelet packet transform is performed on the multi-scale decomposition information to extract the energy mutation intensity coefficients of different frequency bands; Based on the energy mutation intensity coefficients of different frequency bands, construct the pressure frequency domain energy distribution tensor; The spatial gradient direction weight coefficients are subjected to directional enhancement processing to obtain the processed spatial gradient direction weight coefficients. Based on the processed spatial gradient direction weight coefficients, an anisotropic diffusion filtering algorithm is used to generate a gradient direction consistency matrix. The directional weight correction factor is calculated based on the gradient direction consistency matrix and the preset leakage diffusion direction constraint. The pressure frequency domain energy distribution tensor and the direction weight correction factor are input into the dual-channel attention gating module to establish a nonlinear correlation mapping between the pressure frequency domain and the spatial gradient through the cross attention module in the dual-channel attention gating module, thereby generating an initial attention weight matrix. A nonlinear activation operation is performed on the initial attention weight matrix to generate initial values ​​for the time-varying gain factor.

6. The method according to claim 1, characterized in that, The process involves inputting the multidimensional feature matrix into a bidirectional interactive network to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross-attention module in the bidirectional interactive network. A gradient field reconstruction algorithm is then used to spatially propagate the temperature gradient distribution field and the vibration spectrum waveform, generating a fused feature tensor. This includes: Perform time-frequency joint analysis on the pressure fluctuation sequence to generate a pressure fluctuation time-frequency feature matrix; Perform spatial multi-resolution analysis on the gas concentration spectrum to generate a multi-scale feature map of gas concentration; The pressure fluctuation time-frequency feature matrix and the gas concentration multi-scale feature map are input into the cross-attention module to establish the correlation mapping between the pressure frequency domain features and the gas concentration spatial features through the multi-head attention module, thereby generating a pressure-concentration interaction feature matrix. A temperature gradient propagation path map is generated based on the temperature gradient distribution field. The intrinsic mode function components of the vibration spectrum waveform are extracted using the empirical mode decomposition algorithm, and a vibration spectrum energy feature vector is generated based on the intrinsic mode function components. Based on the temperature gradient propagation path diagram and the vibration spectrum energy feature vector, a temperature vibration interaction feature matrix is ​​generated; The pressure-concentration interaction feature matrix and the temperature-vibration interaction feature matrix are subjected to tensor concatenation to generate a fused feature tensor.

7. The method according to claim 1, characterized in that, The process of determining entropy abrupt change regions based on the energy loss vector, and triggering multi-level early warning signals and generating a leak source thermal map and pressure balance parameter set when the high-density region in the leakage probability cloud map spatially overlaps with the entropy abrupt change region within a continuous detection period, includes: Perform a multi-dimensional decomposition operation on the energy loss vector to generate an energy loss feature matrix; Abnormal energy entropy values ​​are identified based on the energy loss feature matrix to generate entropy value mutation regions; Spatial clustering analysis is performed on the leakage probability cloud map to generate high-density regions; When the high-density region and the entropy value mutation region spatially overlap for at least two consecutive detection cycles, a multi-level early warning signal is triggered. Generate a leakage source thermal map corresponding to the high-density area. Based on the leakage source thermal map and the preset pressure balance equation, calculate the pressure compensation parameters of multiple adjacent hydrogen energy storage devices to generate a pressure balance parameter set.

8. A deep learning-based early warning system for leaking hydrogen energy storage equipment, characterized in that, include: The collection module is used to collect multi-source heterogeneous sensor data streams from hydrogen energy storage devices in real time to construct a multi-dimensional feature matrix. The sensor data streams include pressure fluctuation sequences, temperature gradient distribution fields, vibration spectrum waveforms, and gas concentration spectra. The input module is used to input the multidimensional feature matrix into the bidirectional interactive network, so as to establish a nonlinear mapping between the pressure fluctuation sequence and the gas concentration spectrum through the cross attention module in the bidirectional interactive network, and to use the gradient field reconstruction algorithm to perform spatial propagation modeling of the temperature gradient distribution field and the vibration spectrum waveform to generate a fused feature tensor. The output module is used to input the fused feature tensor into the spatiotemporal inference model, so as to use the spatiotemporal inference model to predict the leakage and energy loss of hydrogen energy storage equipment, and output the leakage probability cloud map and energy loss vector. The triggering module is used to determine the entropy value mutation region based on the energy loss vector. When the high-density region in the leakage probability cloud map and the entropy value mutation region spatially overlap within a continuous detection period, a multi-level early warning signal is triggered and a leakage source thermal map and pressure balance parameter set are generated. The entropy value mutation region refers to the region in the energy loss vector where the entropy value changes, indicating abnormal energy fluctuations. The activation module is used to activate the self-regulating system based on the leakage source thermal map and the pressure balance parameter set, so that the self-regulating system dynamically corrects the valve opening gradient of adjacent hydrogen energy storage devices and updates the node weights of the bidirectional interactive network to generate a closed-loop link.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a method for early warning of leakage faults in hydrogen energy storage equipment based on deep learning as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a method for early warning of leakage faults in hydrogen energy storage devices based on deep learning, as described in any one of claims 1 to 7.

Images